Memory is where agent lock-in lives — without it, agents are commoditized

Chase argues that switching model providers has been easy precisely because providers are stateless — “you have to change prompts a little bit, but that’s not that hard.” The moment state is associated with an agent, switching becomes costly, because memory is what makes the agent personalized to a user’s preferences, tone, and usage patterns. Without memory, “your agents are easily replicable by anyone who has access to the same tools.” The article’s personal anecdote — Chase’s internal email assistant was accidentally deleted, and recreating it from the same template produced a much worse experience because the accumulated memory was gone — is the stickiness argument in concrete form.

This reframes the competitive question: the moat isn’t the model weights or even the prompt, it’s the proprietary dataset of interactions and preferences the agent builds up. It compounds Context is the product, not the model and extends Proprietary feedback loops create moats that widen with every interaction by locating the feedback loop specifically in memory state. It also explains why model providers are so incentivized to push memory behind their APIs — see Closed harnesses behind APIs create memory lock-in by design. For vertical builders, this argues that owning the memory layer is a more durable position than owning the model or the UI. Chase extends this in a later talk: Memory defines the agent — a zip of markdown files IS the agent, and portable memory between harnesses is the frontier — memory doesn’t just create lock-in, it defines the agent’s identity. A zip of markdown files (system prompt + skills + tools) IS the agent. This makes portable memory the frontier for agent interoperability. Within memory types, Procedural memory is the highest-impact type of agent memory — it determines what the agent actually does identifies instructions/skills/tools as the highest-impact layer — the one that determines what the agent actually does differently after each correction.